mnn_device.m

基于BP模型的神经网络模型

%~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
%
%  mNN_device(n,beta,msize) - create structure to work with.
%  
%  Parameters: n    - number of neurones
%              beta - initial weights from [-beta,beta]
%              msize - rows and cols of input matrix
%              alpha - inertia factor (if using gradient descent)
%              eta - learning rate (also if using gradient descent)
%              epsilon - stopping criteria (if error < epsilon - exit)
%              epsilon1 - if weight deltas < epsilon1 then exit
%              earlyStop - if error increases more thanEarlyStop iterations then stop 
%
%  Author: Povilas Daniu餴s, [e-mail]
%  http://ai.hacker.lt - lithuanian site about Artificial Intelligence.
%
%  TODO: weighted MSE minimization, maximal likelihood method, multiple
%  activation function support.
%  ----------------------------------------------------------------------

function f = mNN_device(n,msize,alpha,eta,epsilon,epsilon1,earlyStop)

ro = msize(1);   
co = msize(2);
 
for i=1:n+1
    left(i).w = (2*rand(1, ro) - 1);
    right(i).w = (2*rand(co, 1) - 1); 
    d_left(i,:) = zeros(1,ro);  %derivatives
    d_right(i,:) = zeros(co,1);
    dleft(i,:) = zeros(1,ro);   %deltas
    dright(i,:) = zeros(co,1);
end    

f.regressors = n+1;
f.left = left;
f.right = right;
f.d_left = d_left;
f.d_right = d_right;
f.dleft = dleft;
f.dright = dright;
f.d_b = zeros(1,n+1);
f.weights = 2*rand(1,n+1) - 1;
f.dweights = zeros(1,n+1);
f.d = zeros(1,n+1);
f.bias = (2*rand(1,n+1) - 1);
f.dbias = zeros(1,n+1);
f.numparams = n+1 + (n+1)*(ro + co + 1);
f.alpha = alpha;
f.eta = eta;
f.epsilon = epsilon;
f.earlyStop = earlyStop;
f.epsilon1 = epsilon1;
f.name = 'Neural network with matrix inputs';
% -------------------------------------------------------------------------
% -